UIVNAV: Underwater Information-driven Vision-based Navigation via Imitation Learning

TL;DR

UIVNav tackles underwater navigation without localization by learning a domain-invariant policy through imitation on an intermediate representation derived from depth estimation and OOI segmentation. The two-stage approach first builds $I_{DS}$ to suppress domain-specific cues, then trains a policy using $I_{DS}$ to maximize information gain about OOIs while avoiding obstacles; the method generalizes across OOIs (e.g., oysters, rock reefs) and domains (simulation, pool) and outperforms random walk and complete coverage for the same travel distance. Real-time pool experiments with a BlueROV2 demonstrate practical viability. This work advances generalized, localization-free underwater exploration and motivates open-water testing and sensor-fusion enhancements such as sonar.

Abstract

Autonomous navigation in the underwater environment is challenging due to limited visibility, dynamic changes, and the lack of a cost-efficient accurate localization system. We introduce UIVNav, a novel end-to-end underwater navigation solution designed to drive robots over Objects of Interest (OOI) while avoiding obstacles, without relying on localization. UIVNav uses imitation learning and is inspired by the navigation strategies used by human divers who do not rely on localization. UIVNav consists of the following phases: (1) generating an intermediate representation (IR), and (2) training the navigation policy based on human-labeled IR. By training the navigation policy on IR instead of raw data, the second phase is domain-invariant -- the navigation policy does not need to be retrained if the domain or the OOI changes. We show this by deploying the same navigation policy for surveying two different OOIs, oyster and rock reefs, in two different domains, simulation, and a real pool. We compared our method with complete coverage and random walk methods which showed that our method is more efficient in gathering information for OOIs while also avoiding obstacles. The results show that UIVNav chooses to visit the areas with larger area sizes of oysters or rocks with no prior information about the environment or localization. Moreover, a robot using UIVNav compared to complete coverage method surveys on average 36% more oysters when traveling the same distances. We also demonstrate the feasibility of real-time deployment of UIVNavin pool experiments with BlueROV underwater robot for surveying a bed of oyster shells.

Figures (7)

• Figure 1: Example of two trajectories of BlueROV surveying oyster-reef in a pool. On the bottom left is the current frame in the intermediate representation that the robot observes.
• Figure 2: An overview of our approach. UIVNav consists of two phases: Generating Intermediate Representation (gray box) and Data Invariant Navigation Policy (orange box). Image ($I_D$) is fed into the depth estimation network and the segmentation network. The output from the depth estimation network ($I_D$) and the output from the segmentation network ($I_S$) are used to generate the intermediate representation ($I_{DS}$). This is fed into an imitation learning network to train a navigation policy that predicts yaw and pitch changes. Due to training on ($I_{DS}$) the navigation policy is domain invariant and will not require restraining nor labeling with new OOI.
• Figure 3: Example of BlueROV surveying reef of oysters in simulation (translucent mask is applied on the oyster reef for better visualization.)
• Figure 4: Comparing various exploration strategies across four oyster reef patterns-- (a) Grid world, (b) E-shape, (c) Disconnected Paths, (d) Branching Corridor.
• Figure 5: UIVNav surveying reef of rocks OOI with no additional human-expert labeling or training. (a) The current frame of the robot field of view in OysterSim simulation. (b) Intermediate representation of the current frame that the robot uses to make navigation decisions. (c) The complete trajectory over a rock reef.